[Paper Review] Building Damage Detection in Satellite Imagery Using Convolutional Neural Networks

Joseph Z. Xu, Wenhan Lu+3 more|arXiv (Cornell University)|Oct 14, 2019

Remote-Sensing Image Classification11 references116 citations

TL;DR

The paper compares four CNN architectures for detecting damaged buildings in post-disaster satellite imagery, finds twin-tower subtract (TTS) best, and shows cross-region generalization improves with multi-region training and fine-tuning.

ABSTRACT

In all types of disasters, from earthquakes to armed conflicts, aid workers need accurate and timely data such as damage to buildings and population displacement to mount an effective response. Remote sensing provides this data at an unprecedented scale, but extracting operationalizable information from satellite images is slow and labor-intensive. In this work, we use machine learning to automate the detection of building damage in satellite imagery. We compare the performance of four different convolutional neural network models in detecting damaged buildings in the 2010 Haiti earthquake. We also quantify how well the models will generalize to future disasters by training and testing models on different disaster events.

Motivation & Objective

• Motivate automated, scalable damage assessment to accelerate humanitarian response.
• Develop a dataset of pre- and post-disaster satellite image crops labeled for building damage.
• Compare multiple CNN architectures to identify robustness to misalignment and pre/post-disaster variability.
• Assess how models generalize across different disaster regions and events.

Proposed method

• Create a dataset of pre- and post-disaster image crops centered on buildings using a data-generation pipeline.
• Label damaged buildings using UNOSAT assessments and generate negatives via building detections.
• Evaluate four CNN architectures (CC, PO, TTC, TTS) based on how they fuse pre/post-disaster information.
• Use 5-fold cross-validation on the Haiti dataset and AUC as the primary metric.
• Select the best-performing architecture (TTS) for cross-region generalization experiments.

Experimental results

Research questions

• RQ1 Can CNN-based models accurately detect damaged buildings from satellite imagery across disasters?
• RQ2 Which input fusion strategy (combined channels, single post-disaster image, twin-tower with concatenation or subtraction) yields the best performance?
• RQ3 How well do models trained on one region disaster generalize to other regions?
• RQ4 Does augmenting training data with multiple regions or fine-tuning on a small regional set improve cross-region performance?

Key findings

• Twin-tower subtract (TTS) architecture achieved the highest validation AUC of 0.8302 on the Haiti dataset.
• Twin-tower models outperform single-tower models, indicating value in comparing pre-/post-disaster information abstractly.
• The Post-image Only (PO) model outperformed the Concatenated Channel (CC) model, suggesting that feature extraction before fusion helps with misalignment and lighting differences.
• Cross-region experiments show higher AUC when training includes multiple regions, and further gains when fine-tuning with a small regional subset.
• Best cross-region results are obtained when a small amount of regional labeling is used to tune the model, followed by applying it to the larger area.